how distorted food prices discourage a healthy diet€¦ · environment on diets is smaller than...
Post on 27-Jul-2020
6 Views
Preview:
TRANSCRIPT
How Distorted Food Prices Discourage a Healthy Diet
Roberto Pancrazi
University of Warwick
R.Pancrazi@warwick.ac.uk
Thijs van Rens
University of Warwick and Centre for Macroeconomics
J.M.van-Rens@warwick.ac.uk, tvanrens@gmail.com
Marija Vukotic
University of Warwick
M.Vukotic@warwick.ac.uk
February 2020∗
Abstract
We use data on purchase quantities and prices of healthy and unhealthy food to
show that diets are determined at least partially by the “food environment”. We
then estimate a structural model to argue that the effect of the environment is im-
portant: Distorted relative prices lead to underconsumption of fruit and vegetables
in the amount of around 15%, a third of the gap between the recommended and
actual intake. We show that it is possible to remedy these distortions with a policy
of taxes and subsidies that makes all consumers better off.
Keywords: food purchases, food prices, food consumption, retail access, supply ex-
ternalities, food deserts, nutritional inequality, healthy diet, obesity
JEL codes: D12, (E21,) H23, I12, I14, I18, L81
∗We thank Noriko Amano Patino, David Atkins, Diego Comin, Chris Edmond, Jessie Handbury, PaulHansen, Matthias Schundeln and Becca Taylor for helpful comments and suggestions. The empiricalresults in this paper are based on data from The Nielsen Company (US), LLC and marketing databasesprovided by the Kilts Center for Marketing Data Center at The University of Chicago Booth School ofBusiness. The conclusions drawn from the Nielsen data are those of the researchers and do not reflectthe views of Nielsen. Nielsen is not responsible for, had no role in, and was not involved in analyzingand preparing the results reported herein. We are grateful to the University of Warwick Global ResearchPriority in Food for funding the purchase of the Nielsen data.
1
1 Introduction
Diet-related disease leads to more preventable deaths in the US than any other risk
factor. Obesity and overweight by itself was the third-largest preventable cause of deaths
(11%) in 2005, after smoking (24%) and high blood pressure (20%). Including other
diet-related risk factors1 increases that contribution to 28% (Danaei, Ding, Mozaffarian,
Taylor, Rehm, Murray, and Ezzati (2009)). The situation in most other developed
economies is similar.2 This is a relatively new problem, with obesity rates increasing
sharply since 1980 (Ogden, Carroll, Curtin, McDowell, Tabak, and Flegal (2006)), which
is associated with staggering economic as well as human costs (Finkelstein, Trogdon,
Cohen, and Dietz (2009)). What determines our diets? And is there a role for policy to
improve what we eat?
A large literature in public health explores the effect of “obesogenic environments”
(Townshend and Lake (2017)), a term first coined by public health expert Boyd Swinburn
(Pincock (2011)). The environment may matter for obesity because of its effect on
inactivity (opportunities for exercise and active transport) or on diet, through access
to healthy food, advertising, labelling and food prices. We focus on the effect of prices,
which -as economists- we would consider likely to be the most important aspect of the
environmental determinants of diet, and which is also the area where the literature is
most lacking in convincing evidence (Cawley (2015)).3
It seems plausible that prices are least partly responsible for the deteriorating trend
in diet and therefore the obesity epidemic. Aggregate prices for healthy food groups,
in particular fruit and vegetables (Sturm (2005)), have increased relative to prices of
unhealthy food and drink (Wendt and Todd (2011)). The timing in these relative price
changes roughly lines up with the start of the obesity epidemic.4 However, there is a limit
to what we can learn from the time series, and it is almost impossible to disentangle
different possible explanations using aggregate data alone. In this paper, we study
variation in diet across households of different income levels in an attempt to better
understand the determinants of dietary choices, which will also help us to understand
the aggregate change in diets.
1High dietary salt (5%), low dietary omega-3 fatty acids found in seafood (4%), high dietary transfatty acids (4%), low intake of fruits and vegetables (3%), and low dietary poly-unsaturated fatty acids(1%). This excludes alcohol consumption, which is a sizeable risk factor for preventable death as well(3%).
2In the UK, for instance, the chief executive of NHS England warned that “Obesity is the newsmoking”, with obesity and overweight about to overtake smoking as the biggest cause of cancer (BBC(2014)).
3Cawley (2015) summarizes the problem as follows: “A large number of papers have regressed obesityor BMI on local food prices; the limitation of this approach is that variation in local food prices may bedue to differences in demand over geography and changes in demand over time and thus the estimatesof the impact of local food prices on weight may be biased.” (p.248) “Economic research with goodidentification strategies often is unable to reject the null hypothesis of no effect from many possiblecauses, such as food prices, ...” (p.261).
4The obesity epidemic in the US is generally considered to have started around 1980 (Ogden, Carroll,Curtin, McDowell, Tabak, and Flegal (2006)), although BMI among children started to increase at leasta decade earlier (Bell, Rogers, Dietz, Ogden, Schuler, and Popovic (2011)) and some studies have arguedits roots may date back as far as the early 20th century (Komlos and Brabec (2011)).
2
Our objective in this paper is to decompose the determinants of dietary choices
into preferences or demand shifters on the one hand, and the food environment or
supply shifters on the other. If diets are mostly determined by preferences, then the
role for government intervention may be limited. Therefore, we are most interested
in quantifying the role of distortions in the environment that generate inefficiencies in
dietary choices.
We start with a reduced-form analysis of purchase quantities and prices of food
from the Nielsen HomeScan data. We document that relative consumption of fruit and
vegetables, as a proxy for healthy food more generally, increases strongly with household
income as well as with average income in the county where households reside. This
dietary inequality is the variation that we aim to explain. We also show that relative
prices of fruit and vegetables increase with household income but decrease with average
county-level income for the poorer counties. These findings indicate that both demand
and supply factors are important for diets. If differences in diet were driven exclusively
by demand shifters, e.g. preference heterogeneity, then differences across households
would trace out the supply curve, and we would expect to see a positive correlation
between relative consumption and relative prices.
Our next step is to use a structural model to impose just enough structure on the
data to allow us to quantify the relative importance of demand and supply factors for
diets. The crucial element of the model is the retail technology for food. We assume that
there may be fixed costs associated with retailing food, which we think of as the costs
of the store and supply chain. The fixed costs generate a “backward-bending” supply
curve: the more food of a particular type is sold in a county, the cheaper it is, because
the fixed costs can be spread out over a larger amount of sales. This model predicts that
relative prices are inefficiently high for foods with large fixed costs of supply. We make
the model flexible enough to match the data by assuming that preference parameters
may vary exogenously with income. This allows for preferences to determine diets. We
then estimate the model so that the data inform us about the quantitative importance
of the various determinants of dietary inequality.
The identifying assumption we use to separate demand and supply factors is that
households that live and shop in the same location face the same prices. This assumption
allows us to use variation in prices and quantities across households within a county to
estimate heterogeneity in preferences and the relative demand elasticity. We then use
prices and quantities across counties to estimate the fixed costs of supplying food as the
remaining variation that is not explained by preferences. We implement the estimation
using non-linear simulated method of moments on the same data that we used for the
reduced-form analysis.
We find evidence for distortions in the relative price of healthy food, driven by a
large fixed costs in the supply of fruit and vegetables. Our estimates suggest that these
distortions result in a relative price that is about 40% higher and relative consumption
of fruit and vegetables that is about 15% lower than efficient. This effect of the food
3
environment on diets is smaller than the effect of preferences, but it is economically and
statistically significant and large. Moreover, there are many possible other externalities
that our not included in our model, and our estimates should therefore be interpreted
as a lower bound for the role of the food environment in discouraging a healthy diet.
Since dietary inequality due to price distortions in inefficient, there is a role for gov-
ernment intervention to improve diets. We show that, while any remedy for distortions
in food prices will generally affect the distribution across consumers, it is theoretically
possible to design an intervention that replicates relative prices as in the first-best allo-
cation and makes all consumers strictly better off. We then discuss a feasible tax and
subsidy policy that closely approximates this compensated Pareto-improving interven-
tion, and document its welfare gain and distributional effects.
Two recent papers in economics ask questions that are closely related to this study,
use the same data, but reach conclusions that are quite different and seem even directly
opposed to ours. Allcott, Diamond, Dube, Handbury, Rahkovsky, and Schnell (2019)
show that households of different income levels that buy food in the same store never-
theless make different choices over healthy and unhealthy food, illustrating the role of
preferences for diets. They also show that households that move to a healthier environ-
ment for exogenous reasons do not (immediately) change their diets very much, casting
doubt on the idea that there is an important role for the environment. Amano-Patino
(2020) estimates a demand system for food products on household panel data and sim-
ilarly finds that the role of prices in explaining differences in diet across households of
different income levels is very limited. We confirm the findings in these studies that the
effect of distortions in the food environment on dietary inequality is small. However, our
structural model estimates also allow us to show that these cross-sectional differences in
diet nevertheless indicate an important effect of distorted food prices on diets. In the
terminology of a randomized experiment: we use the (small) differences between poor
households (the treatment group) and rich households (the control group) to estimate
the effect of food prices on diets. These estimates then indicate that the effect of prices is
large, but affects both groups of consumers, with only small differences in the treatment
between treatment and control group.
The remainder of this paper is organized as follows. In section 2 below, we introduce
the Nielsen HomeScan data and document some patterns in dietary inequality that lead
us to conclude that healthy or unhealthy dietary choices are not determined exclusively
by preferences. To quantity the role of the food environment, we need to put some
structure on the data. Section 3 presents a simple and flexible structural model of con-
sumers’ choices between healthy and unhealthy food. We then discuss how we identify
and estimate this model in section 4. In section 5, we turn to our results. We show
that distorted relative food prices due to high fixed costs in the supply lead to inefficient
underconsumption of fruit and vegetables by 14 to 15.5%. Fixing this distortion would
close between a quarter and almost all of the gap between actual and recommended
intakes of fruit and vegetables. We also show that it is theoretically possible for the
4
government to intervene in such a way that makes consumers of all income levels and in
all locations strictly better off, and that such an intervention can be closely replicated
by a subsidy on fruit and vegetable consumption funded by an increase in progressive
income taxes.
2 Purchases of healthy and unhealthy food
We start this paper by documenting some patterns in food purchases by income. First,
we describe the data from the Nielsen HomeScan consumer panel, which we use to
measure food purchases. Then, we argue that one of the most striking differences in food
purchases between rich and poor that impacts a healthy diet is the underconsumption of
fruit and vegetables, which is much stronger for poor households. Finally, we document
that the comovement of quantities and prices of fruit and vegetables purchases in counties
of different income levels indicates that dietary inequality cannot be driven exclusively
by preferences or other demand factors.
2.1 Nielsen HomeScan Data
The HomeScan consumer panel, collected by Nielsen and distributed for academic use
by the Kilts Center for Marketing at Chicago Booth,5 is a dataset containing detailed
information about quantities and prices of food purchases by consumers. Participating
households are given a bar code scanner, and are instructed to scan their food purchases
at home and enter quantity purchased and expenditure for each item, so that unit prices
can be obtained simply by dividing expenditure by quantity. Basic demographic infor-
mation about the household is recorded as well. Importantly for this study, demographic
information includes household income and place of residence (ZIP code).
Purchase information is available for about 60, 000 households. Each household
records on average 172 shopping trips, and on each given trip records on average 7
product purchases with distinct Universal Product Code (UPC). The data are highly
disaggregated by product, because different varieties, different brands and different pack-
age sizes are all coded as different UPCs, resulting in 1.7 million distinct food products.
The complete dataset has around 50 million observations (product purchases) per year,
for ten years from 2004 to 2014.
A problem arises with random weight-items, foods that consumers purchase in any
quantity they desire and are charged for by weight or quantity. This problem affects
mostly fresh produce, but also unprepared meat, poultry and seafood; bread and baked
goods; deli counter items like cheese, deli meat, prepared salads and foods; bread and
baked goods; coffee; and dried vegetables and grains. In the dataset, these purchases are
recorded as “magnet data” with a generated UPC that is much more aggregated than
an actual UPC. In the first years of the sample, magnet data are only recorded as such.
5https://www.chicagobooth.edu/research/kilts/datasets/nielsen
5
From 2007, the “product group” of these magnet data is also recorded.6 Product groups
are 64 groups of similar UPC-level products, defined by Nielsen, which each constitute
up to 4.1% of the average household’s food expenditures. Examples of product groups,
ordered by expenditure share, are “bread and baked goods”, “snacks”, “packaged meats
- deli”, “fresh produce”, “cheese”, “prepared foods - frozen”, “milk”, “carbonated bever-
ages”, “candy” and “juice drinks - canned, bottled”. Since fresh produce is particularly
important for this study, we only use data for the 2007-2010 period, and we aggregate
products to the product group level.
We use a two-step procedure to aggregate product data to product groups as fol-
lows. First, we calculate household-level aggregate quantities for product groups as the
weighted sum of the quantities of all products in that group, using as weights the average
price of each product across households. Aggregating prices similarly is not possible,
because the household-specific prices are not observed for products that a particular
household does not purchase. Therefore, as the second step, we calculate household-
level aggregate prices for product groups by dividing household-level expenditures by
the household-level aggregate quantity. Apart from dealing with zero expenditures, this
procedure has the additional advantage that total expenditures are preserved with ag-
gregation. We normalize all aggregate prices to have an average of one across households,
so that all quantities are expressed as numbers of baskets of all products in a product
group.
Household income is provided in the Nielsen data. We add average income in the
location of residence for each household by matching their ZIP code to income data from
the Census,7 and then aggregating to the county level.8 Household income is coded as a
categorical variable, which we make continuous by assigning housholds an income level
corresponding to the midpoint of the income range represented by the income category
they are in. Household income then ranges from from $2500 to $120,000 per year in 16
categories.
2.2 Dietary inequality
We start by documenting how purchases of different types of food vary with income.
To document the relation between food purchases and income, we first project purchase
quantities and prices on a 10-th order polynomial in household and location income, and
then aggregate to 5%-tile bins in these variables. In some specifications, we also correct
for household composition and demographics, by including second-order polynomials
in household size, age and years of schooling (normalized at the sample mean), and
dummies for whether 2 adults are present in the household, and for household heads
6From 2010, the product group of magnet data is directly recorded in the data. Over the 2007-2010period, this information can be retrieved from the generated UPC for these purchases.
7https://www.psc.isr.umich.edu/dis/census/Features/tract2zip/8The number of ZIP codes is too large compared to the number of households in our dataset and
results in too few observations in each ZIP code to be able to document the relation between foodpurchases and income at this level. There are 3007 counties in the US, so that we have on average 20households in a county.
6
that are female, African American, Asian, other race, hispanic, part-time employed and
non-employed in the regression. Finally, we correct income (as well as food purchases)
for inflation and real income growth by normalizing the mean across households in each
year to equal the mean in 2010.
For each product group, we plot the quantity and the price at which foods in this
group were purchased against household and county-level income in 5%-tile bins. These
results are presented in appendix A. Assuming that food consumption closely tracks
food purchases, these graphs reflect differences in diet across households and locations
of different income levels. We analyze these dietary differences with a focus on what
may explain inequalities in health, particularly obesity.
One pattern in dietary inequality stands out, because it is both striking and likely
to affect health outcomes. Poorer households consume much less fresh produce than
richer households do. We argue that this difference reflects a broader difference in
the healthiness of the diet of poorer versus richer families. We show that the higher
underconsumption in poor households of fruit and vegetables does not depend on the
definition of fruit and vegetables. In figure 1, we show the consumption of fruit and
vegetables against household income, as a fraction of total food consumption. The diet
of the poorest 5% of households contains around 40% less fruit and vegetables than the
diet of the richest 5%. This difference is large compared to the overall underconsumption
of fruit and vegetables relative to the recommended intake of 40 − 60% above average
current intakes (U.S. Department of Health and Human Services and U.S. Department
of Agriculture (2020), Figures 2.3 and 2.4). We show that this finding is robust to four
different definitions of fruit and vegetables: (1) only fresh produce, as in the figure, (2)
also including frozen vegetables,9 (3) also including nuts, dried fruit, dried vegetables
and grains, and (4) also including canned vegetables, canned fruit, and pickles, olives
and relish.
9It is not possible to include frozen fruits, because these are coded in the same product group asdesserts.
7
Figure 1: Relative Consumption vs Individual Income
Note: This figure plots consumption of fruit and vegetables against household income, as a fraction
of total food consumption, against household income. Each point is a 5-percentile bin. The shaded
area represents a two standard-error confidence band.
Vegetables and fruits (and whole grains) are the most important two (three) of the
six components of a healthy eating pattern (U.S. Department of Health and Human
Services and U.S. Department of Agriculture (2020), chapter 1, p.15) and also the food
groups, for which current intakes are furthest below recommended levels. The aim of
this paper is to explore what causes this underconsumption. We will heavily rely on the
cross-sectional differences in fruit and vegetable consumption documented in figure 1 as
a source of identification, and as a by-product of our analysis, we will have some results
on the determinants of dietary inequality as well. We think of the share of fruit and
vegetables in a diet as a good summary measure of a healthy diet. We experimented
with various other definitions of healthy and unhealthy food, and always found very
similar patterns with income as for the relative consumption of fruit and vegetables.
2.3 Demand and supply determinants of healthy diets
What drives the difference in diet between poor and rich households? Clearly, there is
a role for demand factors. Innate preference over food differ across people, and these
preferences may be correlated with other characteristics and income. More generally,
“preferences” may be shaped by education, parental income and education, and social
economic status, all of which are correlated with income as well. There may also be
differences across consumers in the extent that they are aware of and care about the
health effects of their diet. On the other hand, it is equally clear that supply factors
may matter too. Much has been written about obesogenic environments, which may
feature limited access to healthy food, abundant advertising and marketing for unhealthy
food, and misleading food labels. One way, arguably the most important way, in which
8
the food environment affects dietary choices is through relative prices of healthy and
unhealthy food options. Recently, Allcott, Diamond, Dube, Handbury, Rahkovsky, and
Schnell (2019) make a powerful argument that preferences are an important determinant
of diets, and cast doubt on the relevance of the environment.
Comparing the prices paid for fruit and vegetables by income levels with quantities
purchased gives some insight into the relative importance of demand and supply factors.
These prices are shown in figure 2. Comparing relative prices of fruit and vegetables
in this figure to the relative quantities in figure 2 above seems to indicate that supply
factors are not very important: richer households not only consume a higher fraction of
their diet in the form of fruit and vegetables, but they also pay relative more for these
healthy foods. This suggest that their demand is higher, thus driving up the prices.
Figure 2: Relative Prices vs Individual Income
Note: This figure plots the relative price of fruit and vegetables against household income. Each point
is a 5-percentile bin. The shaded area represents a two standard-error confidence band.
But household-level purchase prices are not a good measure of market prices. There is
a wide variety of products within the “fruit and vegetables” category, and the household-
level price of the basket of fruit and vegetables will be determined in large part by
the household’s choice over which products from that basket to purchase. It seems
reasonable to expect that richer households will tend to buy higher-quality or more
desirable, more expensive types of fruit and vegetables, which may explain why these
households end up paying higher prices for this category of food.
A better measure of the food environment are the relative prices of healthy food in
a location. Figures 3a and 3b show the relative quantity of fruit and vegetables pur-
chased and the relative price of fruit and vegetables against income across counties. The
inequality in diet between poor and rich households carries over to inequality in diet
between poor and rich counties. The underconsumption of fruit and vegetables in the
9
poorest 5% of counties is about 25% larger than in the richest 5%, compared to a differ-
ence of 40% between the poorest and richest 5% of households. However, the differences
in income are also smaller across counties, and the gradient of underconsumption of fruit
and vegetables with income is actually higher across counties than across households.
Figure 3: Relative Consumption and Prices across locations
(a) Relative Consumption (b) Relative Prices
Note: The left panel plots the consumption of fruit and vegetables against income across counties, as
a fraction of total food consumption; the right panel plots the relative price of fruit and vegetables
against income across counties. The shaded areas represent a two standard-error confidence band.
Each point is a 5-percentile bin.
Although dietary inequality across counties looks similar or even stronger than across
households, the picture for relative prices now looks quite different. For the richest half
of counties, differences in diet still seem to be driven largely by differences in preferences,
as richer counties face less favorable relative prices of fruit and vegetables. However,
across poorer counties, the poorer the county, the higher the relative price of fruit and
vegetables. It is not possible to reconcile this negative correlation between relative
quantities and prices of healthy food with an standard upward-sloping supply curve
that is traced out by differences in demand factors.
We conclude that there must be a role for the environment in determining diets, at
least in poorer counties. Without further structure on the data, this is as far as we can
go. We now turn to a structural model that will allow us, with minimal assumptions,
to quantify the role of supply factors in determining diets.
3 A simple model of dietary choice
In this section we describe our model of dietary choice. This model is deliberately
simple. The aim is to impose just enough structure on the data to allow us to do welfare
analysis and counterfactuals decomposing dietary inequality across income levels into
a part that is driven by demand factors or preferences, broadly defined, and a part
that is determined by supply factors or the food environment. We therefore only model
10
consumers demand and retailers pricing decisions, and in both cases we allow for a large
part of the decisions that is not captured by the model through exogenous parameter
heterogeneity.
3.1 Consumers
Households, indexed by i ∈ N and living in location j ∈ K, differ from each other in
their income yij . The distributions of household income yij is exogenous. The model is
static and is assumed to hold in all time periods, so we suppress time subscripts.
Households spend an amount m (yij) of their income yij on food. Food expenditure
is exogenous, and will be chosen to match the data. They choose how to allocate their
total food expenditure between consumption of healthy food CH,ij and unhealthy food
CU,ij , in order to maximize their utility function over food.
u (CH,ij , CU,ij) =[α (yij) (CH,ij)
ε−1ε + (1− α (yij)) (CU,ij)
ε−1ε
] εε−1
(1)
The parameter α (yij) allows for preferences for healthy food to be correlated with income
yij . This non-homotheticity of the utility function is one way, by which preferences may
affect diets. The only real restriction on the variation of diets with income is that
the elasticity of substitution between healthy and unhealthy food ε is assumed to be
constant across households of different income levels.
Consumers maximize their utility u (CH,ij , CU,ij) subject to a standard budget con-
straint,
PH,ijCH,ij + PU,ijCU,ij = m (yij) (2)
where PH,ij and PU,ij are the prices of healthy and unhealthy food, which the household
takes as given. These prices may vary across households to reflect that different house-
holds live and shop in different locations j with potentially different price levels, but
also to allow for the possibility that richer households will buy food of higher quality.
This is the second way, by which preferences may affect diet.
The first-order conditions give the relative consumption of healthy over unhealthy
food as a function of preference parameters and prices.
CH,ijCU,ij
=
[1− α (yij)
α (yij)
PH,ijPU,ij
]−ε(3)
This is a (relative) demand equation, where ε is the price elasticity of demand. A
household consumes a healthier diet if their preference for healthy food α (yij) is stronger
or if healthy food is cheaper relative to unhealthy food.
3.2 Food retailers
Food, both healthy and unhealthy, is supplied by retailers. Retailers buy food from a
wholesaler, so that the marginal cost of healthy and unhealthy food of the same quality
11
is the same in all locations. However, the marginal costs differs with the quality of the
food. To the extent that some households purchase higher quality food than other, the
marginal cost of supplying this food will be higher too. We model this by assuming that
the marginal costs κH (yij) and κH (yij) vary exogenously with household income yij .
In addition to their marginal costs, retailers face a fixed cost of supplying healthy
food KH or unhealthy food KU in a particular location j, which may depend on the
location, so that KH (yj) and KU (yj) may vary exogenously with location income the
average income level of household in the location yj . This dependence reflects the fact
that it may be more expensive to run a store or operate a supply line in an area where
real estate is more expensive and traffic more congested.
We assume there is free entry of food stores in each location. The fixed costs generate
increasing returns to food supply, so that in equilibrium there will be a single store for
each food type in each location. However, the threat of entry implies that the local
monopoly is only sustainable if retailers make zero profits. This assumption is consistent
with the evidence in Altomonte, Barattieri, and Basu (2015), who show that firms price
as a markup over average rather than marginal costs.
A local sustainable monopoly with zero profits implies that prices are set just high
enough to recover fixed costs,
PH,ij =KH (yj)
CH,j+ κH (yij) (4)
PU,ij =KU (yj)
CU,j+ κU (yij) (5)
where CH,j =∑
i∈j CH,ij and CU,j =∑
i∈j CU,ij denote the total consumption in location
j of healthy and unhealthy food, respectively. Note that these pricing equations imply a
“backward bending” supply curve: the more food is sold, the lower the price the retailer
will charge for it.
Two observations are important for future reference. First, the fixed costs associated
with food supply introduce an externality: the more food of a certain type a household
purchases, the cheaper this type of food becomes for other households in the same
location that shop in the same store. This externality is the distortion in the food
environment that will generate inefficiencies in equilibrium diets. Second, although
prices are household-specific, only aggregate demand affects prices (by lowering the
impact of fixed costs on the average cost per unit sold). This property of the model will
be key to our identification strategy, see section 4.1 below.
3.3 Distortions in the food environment
The equilibrium of our model is defined as follows.
Definition 1 An equilibrium in our model is an allocation of quantities {CH,ij , CU,ij}and a set of prices {PH,ij , PU,ij}, ∀i ∈ N and j ∈ K, such that, for a given distribution
12
of income over households and locations, yij for all i ∈ N and j ∈ K:
(i) each household i maximizes its utility, so that equations (2) and (3) are satisfied
∀i;
(ii) each retailer in any of the two sectors set prices satisfying (4) and (5) ∀i, k;
(iii) aggregation identities CH,j =∑
i∈j CH,ij and CU,j =∑
i∈j CU,ij are satisfied for
all locations j ∈ K.
Hence, the equilibrium is a well defined system of 4×N non-linear equations in 4×Nunknowns.
The equilibrium of our very simple model has some interesting properties. In this
paper, we largely ignore these and primarily use the model to quantity the sources of
dietary inequality. However, in order to build some intuition for the model, we briefly
describe some of its predictions here intuitively, with more detail provided in appendix
B.
First, the model allows for a role of preferences as well as supply factors as drivers
of dietary inequality. Preferences broadly defined, generate differences in diet across
households of different income levels because of differences in the preference for healthy
food α (yij) or differences in the preference for quality κ (yij), which leads to differences in
prices. Inequality in diets that is driven by these two channels are efficient in the model,
and there is no role for government intervention.10 But the food environment matters as
well. Households that live in a location where there is low demand for healthy food will
face higher relative prices for healthy food, because the fixed costs have to be distributed
over a smaller base of aggregate sales of healthy food. Inequality in diets originating
from this effect is inefficient. The externality is a monopoly distortion, which creates
a coordination problem among consumers. If the government were able to coordinate
all households in a location to buy more healthy food, then these households would all
benefit from lower prices for healthy food.
The second model prediction we briefly highlight here, is that for sufficiently high
fixed costs the market for a particular type of food, e.g. healthy food, may completely
collapse. This happens when demand for healthy food is low, leading to high prices.
High relative prices of healthy food lead households to further substitute away from
healthy food, exacerbating the effect on prices. Eventually the price of healthy food
exceeds the monetized utility from the first unit of healthy food and in equilibrium no
healthy food is being demanded or supplied. Thus, the model predicts the emergence of
“food deserts” (Algert, Agrawal, and Lewis (2006), Powell, Slater, Mirtcheva, Bao, and
Chaloupka (2007), Larson, Story, and Nelson (2009), Sharkey, Horel, and Dean (2010)).
In this paper, we assume that markets for both types of food will remain open at all
times and do not exploit this interesting prediction, leaving it for future research.
10Of course, many possible externalities are missing from our simple model. For instance, if consumersdo not internalize the cost of health care for preventable disease caused by their diet, then this wouldjustify government intervention.
13
Third, the model gives an interpretation of the deteriorating trend in healthy diets
and the associated obesity epidemic. Distribution costs are an important component
of the cost of food (Kelsey (1989)) and technological progress in food “processing” has
dramatically reduced those costs (Schlosser (2013)). Unfortunately, processing makes
food unhealth, by removing spoilable nutrients (refining), adding chemical flavoring or
other additives and other processes like hydrogen-peroxide sterilization (Cutler, Glaeser,
and Shapiro (2003), Lustig (2012)). As a result, technological progress has been biased
toward unhealthy food, which may explain why relative prices of healthy food have risen
and relative consumption of healthy food has declined. Our model suggests that this
deterioration in diet as a result of technological progress may have reduced welfare.
4 Estimation results
We now proceed to estimate the model using the data described in section 2 and trying
to match the patterns in dietary inequality documented there. After discussing identifi-
cation, estimation and model fit, we briefly consider what we learn from the parameter
estimates. Then, in section 5, we use the estimated model for counterfactual analysis.
4.1 Identification
The objective of the estimation of the model is to find the demand shifters α (yij) and
supply shifters κH (yij) and κU (yij), which together describe the contribution of pref-
erences to differences in diet between households of different income levels, as well as
the fixed costs KH (yj) and KU (yj), which give rise to dietary inequality due to sup-
ply distortions, and other structural parameters like the demand elasticity ε, for which
the equilibrium allocation in our model -given the observed distribution of income over
households and counties in the data- replicates the observed patterns in food consump-
tion and prices as closely as possible. Estimating these parameters runs into the usual
identification problem in a simultaneous supply-demand system.
Our identifying assumption is that, whereas individual household demand depends
on household-specific prices, prices in the supply curve do not depend on individual
household demand, but only on aggregate consumption in the location, see equations
(4) and (5). This assumption allows us to solve the identification problem by using
variation in quantities and prices both across households within locations and across
locations.
Taking within-location deviations of the supply equations (4) and (5) gives PH,ij =
κH (yij) and PU,ij = κU (yij), where a hat over a variable denotes within-location devia-
tions, i.e. Xij = Xij−Xj . These equations immediately identify the income dependence
of the variable costs or preferences over quality. Moreover, since variation in prices across
households within locations is exogenous to household-level demand, we can identify the
relative demand equation in within-location deviations from an ordinary least squares
regression of the log of relative quantities of healthy over unhealthy food on the log of
14
relative prices and income.
log
(CH,ijCU,ij
)= −ε
log
(PH,ijPU,ij
)− ε
log
(1− α (yij)
α (yij)
)(6)
This regression identifies the elasticity of substitution between healthy and unhealthy
food ε, as well as the income dependence of preferences for healthy food α (yij).
The intuition for our identification strategy is that households that live and therefore
shop in the same location face the same prices. Therefore any differences in the amount
of healthy and unhealthy food they purchase, and the prices they pay for these foods,
must be due to their preferences. This idea is very similar to the identification strategy
in Allcott, Diamond, Dube, Handbury, Rahkovsky, and Schnell (2019). However, we
extend their strategy in two ways. First, we disentangle preferences for healthy food
from preferences for quality. If we observe different households paying different prices,
we attribute that to differences in their preferences over quality. If we observe differences
in diet given the same prices, we attribute that to differences in preferences over healthy
versus unhealthy food. Second, having the structure provided by a structural model
allows us to go a step further: having identified the role of preferences from variation
within locations, we can then use the variation across locations to identify the role of
exogenous price shifters as well.
The supply and demand system described by equations (2), (3), (4) and (5) implies a
reduced-form relation of consumption of healthy food CH,ij , consumption of unhealthy
food CU,ij , and prices PH,ij and PU,ij of healthy and unhealth food, respectively, with in-
come. These relations depend on the income dependence of preferences α (yj), marginal
costs κH (yj) and κU (yj), and fixed costs KH (yj) and KU (yj). Since α (yj), κH (yj)
and κU (yj) follow directly from the within-location estimates of α (yij), κH (yij) and
κU (yij), the estimated reduced form allows us to identify the level and the income
dependence of the fixed costs, see appendix C for details.
4.2 Estimation and model fit
We implement the estimation of the model through non-linear simulated method of
moments. We find the model parameters that minimize the distance between model
and data for CH,ij , CU,ij , PH,ij and PU,ij for all levels of income. We summarize the
information in the data as the average value of these variables in 5%-tile bins for income,
both across households in deviation from the location average, and across locations.
This gives 160 moments to match (20 income quantiles for four variables, each both
within and across locations). To faciliate efficient estimation, we approximate the income
dependence of the parameters α (y), κH (y), κU (y), KH (y) and KU (y) as fourth-order
polynomials, which gives us a total of 26 parameters to estimate.
We weigh moments by the inverse of their variance, which we obtain directly from
the microdata. This solves the problem that different moments are in different units,
and also ensures the estimation is (approximately) efficient as in efficient GMM. Since
15
each moment is an average of a variable over a group of households in the microdata, the
variance of the moments is simply the variance of that variable divided by the number
of households in that group. Further details on the estimation procedure are given in
appendix D.
Figure 4 shows the model fit. The model is flexible enough to match almost perfectly
the variation of consumption of healthy food (fruit and vegetables), consumption of
unhealthy food (all other food) and their prices with income, both household income
in deviation from the county average, and county income. The only set of moments
that is matched less well is the steep increase of aggregate consumption of fruit and
vegetables with county income. The difference between the richest 5% of counties and
the poorest 5% is about 27 log points in the data, whereas in the estimated model this
difference is “only” 20 log points. The model could replicate this difference with the
income dependence of the preference for healthy food α (y). However, this parameter
also needs to match the gradient of fruit and vegetable consumption with income within a
county, which is slightly lower. Alternatively, the model could replicate the difference in
consumption of fruit and vegetables between rich and poor counties with a lower level or a
negative income dependence of the fixed costs of healthy food supply KH (y). However,
this would also affect the variation of the price of fruit and vegetables with county
income. The estimation procedure gives more weight to the variation in consumption
across households within counties than to its variation across counties, because the
former has tighther standard errors because of the much larger number of observations
in the underlying microdata. It gives more weight to variation in prices than variation in
consumption across counties, because prices have tighter standard errors because there
is less dispersion in prices than in consmption within counties.
16
Figure 4: Model and Data
Note: This figure displays the model fit. The left column plots consumption levels and the right
column plots price levels. The first row display consumption and prices of fruit and vegetable against
income in deviations from the county mean. The second row display consumption and prices of all
other food against income in deviations from the county mean. The third row display consumption
and prices of fruit and vegetable against county income. The forth row display consumption and
prices of all other food against county income. The purple dotted line plots the data moments with
the two standard error confidence bands displayed as the shaded area, while the blue solid line plots
the model moments. Each point is a 5-percentile bin.
The model fit is comfortably good enough that we can use the estimated model
for counterfactual analysis. The only set of moments, for which the estimated model
systematically and significantly differs from the data is still matched reasonably well.
Importantly, since differences in diet are due to differences within than across counties,
the relation of the relative consumption of fruit and vegetables with individual income,
as in figure 1, is matched almost perfectly, see figure 5.
17
Figure 5: Relative Consumption and Prices vs Individual Income
(a) Relative Consumption (b) Relative Prices
Note: The left panel plots the consumption of fruit and vegetables against household income,
as a fraction of total food consumption; the right panel plots the relative price of fruit and
vegetables against household income. The purple dotted line plots the data moments with
the two standard error confidence band displayed as the shaded area, while the blue solid line
plots the model moments. Each point is a 5-percentile bin.
4.3 Preference parameters
The estimates for the preference over food quality κH (y) and κU (y), and the preference
for healthy food α (y) are given in figure 6. These estimates are best understood by
considering their implications for consumption and prices.
Figure 6: Preference functions
(a) for Food quality
0 50 100 150
Individual Income (in 1000$)
1
1.01
1.02
1.03
1.04
1.05
1.06
1.07
1.08
1.09
Ma
rgin
al C
ost
fun
ctio
n
H(y)
U(y)
(b) for Healthy food
0 50 100 150
Individual Income (in 1000$)
0.6
0.8
1
1.2
1.4
1.6
1.8
Pre
fere
nce function
10-3
(y)
Note: The left panel shows the estimated preference for food quality as a function of household income.
The solid blue line displays preferences for quality of healthy food, κH(y), and the dashed black line
displays preferences for quality of other food, κU (y). The right panel shows the preference for healthy
food in the utility function, α(y).
The estimated preference over healthy food quality corresponds almost one to one
to the price of fruit and vegetables in deviation from the county average, see the top
18
right-hand-side panel in figure 5. Richer households are willing to pay higher prices for
quality fruit and vegetables, and this relation is monotonic and approximately linear.
Households that are twice as rich as (100% richer than) the county average, pay on
average slightly over 2% more for an equal quantity of fruit and vegetables. The same
households pay on average around 1.5% more for other foods (see the second-row right-
hand-side panel in the figure), indicating that the estimated preference for quality over
unhealthy and healthy food is very similar.
The estimated preference for healthy food is determined primarily by the consump-
tion of fruit and vegetables in deviation from the county average, see the top left-
hand-side panel in figure 5. Household demand for fruit and vegetables is affected by
household-specific prices as well, but this effect is small.11 A household that is twice as
rich as the county average buys around 25% more fruit and vegetables than the county
average. This difference is much larger than the difference of around 10% in consumption
of other food. Preferences clearly have a role to play in explaining why richer households
have healthier diets than poorer families.
Relative prices and quantities are closely linked in our model through the elasticity
of substitution between healthy and unhealthy food ε, which is the elasticity of relative
demand for healthy food with respect to its relative price. We estimate ε = 0.44. This
value, which we identify from the comovement of relative prices and relative quantities
within counties, is in line with the literature on price elasticities of demand for food,
which reports elasticities for fruit between 0.41 and 0.98, and for vegetables between
0.44 and 0.71 (Andreyeva, Long, and Brownell (2010)).
4.4 Fixed costs and supply externalities
The estimated fixed costs in the supply of healthy and unhealthy food are shown in
figure 7. The first panel 7a shows the estimated values of the fixed costs and how they
vary with county income. Both fixed costs increase with average income in the county,
as we would expect because retail prices and congestion would tend to make running a
store in richer, more urban counties more costly.
The level of the fixed costs in healthy and unhealthy food cannot be easily compared,
because total expenditures on the two types of food (fruit and vegetables versus all other
food) are very different. Therefore, the second panel 7b in the figure plots the fixed costs
again, but as a share of the prices of healthy and unhealthy food, and against household
income rather than average income in a county. The main finding is that the fixed costs
for supplying fruit and vegetables is much higher than the fixed costs for supplying
other food. For fruit and vegetables, fixed costs account for 20 to 43% of the price,
whereas that same percentage is negligible (less than 1%) for other foods. The second
observation is that while fixed costs are higher in richer counties, they account for a
11For two reasons. First, the higher price of fruit and vegetables is partially offset by higher prices forother foods, due to a similar preference for quality over all foods. Second, given the estimated relativedemand elasticity, the effect of variation in relative prices on relative quantities is small compared tothe observed variation in relative quantities.
19
Figure 7: Fixed Costs
(a) Estimate
30 40 50 60 70 80 90 100
Average County Income (in 1000$)
0
5
10
15
20
25
30
35
40
Fix
Co
st
fun
ctio
n
KH
(y)
KU
(y)
(b) Share
0 50 100 150
Individual Income (in 1000$)
0
5
10
15
20
25
30
35
40
45
Sh
are
Fix
ed
Co
st
(pe
rce
nt)
FVOther food
Note: The left panel displays the estimated fixed cost as a function of county income. The solid blueline displays fixed cost in the healthy food sector, KH(y), and the dashed black line displays fixedcost in the other food sector, KU (y). The right panel displays the share of the price due to fixed cost,in percent, and is computed as (Kz (yj) /Cz,j) / Pz,ij , with z ∈ {H,U}.
larger share of prices for poorer households. This is because poorer households purchase
less food, and less fruit and vegetables in particular, which means that the fixed costs
need to be spread out over a larger amount of sales.
The finding that fixed costs are much higher for fruit and variables than for other
food is important, because it implies that the market for healthy food is much more
distorted than that for unhealthy food. Fruit and vegetable prices are in (large) part
the result of a monopoly distortion that drives up prices where demand is relatively low,
i.e. in poorer counties. The same is not true for other food, which is priced at very close
to marginal cost. This result is what will drive our counterfactual analysis in section 5
below.
Since the estimates of the fixed costs are crucial to our results, it is important to
understand exactly what identifies these estimates. The fixed costs are identified from
the variation in consumption and prices of healthy and unhealthy food across counties, in
particular from the U-shape in the relation of the (relative) price of fruit and vegetables
with average income in the county, see figure 3b. That means that they must help
the model to match both the negative income dependence of prices across counties at
low levels of income, and their positive income dependence at high income levels (over
and above the effect of preferences over quality, which explain some but not all of the
increase in prices with county-level income for richer counties).
The positive income dependence for richer counties is matched simply because fixed
costs increase with county income. The negative income dependence for poorer counties
is matched with the level of the fixed costs. The same level of fixed costs makes up a
larger share of the price in poorer counties, because demand is lower there, so that the
fixed cost needs to be distributed over a lower level of sales. This pushes up the relative
price of fruit in vegetables more in poorer counties. It is encouraging for our story that
20
it is possible to match the U-shape in fruit and vegetable prices across counties with
estimated fixed costs that are monotonically increasing in county-level income, even
though we do not impose this constraint (and that marginal costs are always estimated
to be positive, which we do not impose either).
5 Determinants of an unhealthy diet
We developed a structural model that we argue provides a plausible description of the
variation in consumption and prices of healthy and unhealthy food across households
and locations. We then estimated this model and showed that it fits the data well,
and that the parameter estimates seem reasonable and in line with previous estimates
insofar they can be compared. We now use this model for counterfactual analysis. In
particular, we ask the questions how much diet is affected by the type of distortions in
food prices that we consider in this paper, and what the government can do to fix these.
5.1 Distortions in food prices
We start by comparing the relative consumption of fruit and vegetables in equilibrium to
the first-best allocation. A social planner chooses the allocation of healthy and unhealthy
food for all consumers to maximize the weighted sum of their utility as in (1), using
Pareto weights θij to compare utility across consumers.
max{CH,ij ,CU,ij}i∈N
∑i∈N
θiju (CH,ij , CU,ij) (7)
Retailers make zero profits in equilibrium, so that we do not have to worry about
how profits are redistributed to consumers. The planner is constrained by the retail
technology, the distribution of consumers i ∈ N over locations j ∈ K, and the total
amount of food expenditure m (yij), and must respect the aggregate resource constraint.∑i∈N
[κH (yij)CH,ij + κU (yij)CU,ij ] +∑j∈K
[KH (yj) +KU (yj)] =∑i∈N
m (yij) (8)
In the efficient allocation, the relative consumption of healthy food depends only on
preference and marginal costs, which depend on preferences over quality.
CH,ijCU,ij
=
[1− α (yij)
α (yij)
κH (yij)
κU (yij)
]−ε(9)
Comparing the relative consumption in equilibrium (3) to the efficient allocation (9)
makes clear that equilibrium diets are distorted because prices do not reflect marginal
costs, given preferences for quality. Notice that the efficient relative consumption, and
therefore the distortions, do not depend on the Pareto weights.
The efficient allocation may also differ from the equilibrium in the distribution of
food over consumers: the social planner may want to redistribute food with respect
21
to the equilibrium consumption allocation. The efficient amount of redistribution will
depend crucially on the planner’s Pareto weights. For reasons of exposition, we will
start with the case of equal Pareto weights, θij = 1 for all i ∈ N , which will involve a
good amount of redistribution. However, we will focus on the case where Pareto weights
are chosen such that there is no redistribution at all. We do this in order to focus on the
distortions in relative food prices, in search of a government intervention that knows no
losers but only winners. The question what is the optimal amount of redistribution is
unrelated to the food environment and therefore outside the scope of this paper.
To find the efficient allocation without redistribution, we have to take a stance on
what “no redistribution” means. Clearly, the distributions of healthy and unhealthy food
over consumers will be different in the efficient allocation compared to the equilibrium,
and it is not straightforward to aggregate consumption of both types of food to total
food expenditures, because prices are not defined in the efficient allocation. We solve
for the no-redistribution efficient allocation by imposing two additional constraints on
the social planner’s problem. First, we assume that the planner does not move resources
across locations, so that the fixed costs of retailing healthy and unhealthy food in each
location must be paid for by consumers in that location. Second, we assume the planner
must fund the fixed costs by levying a proportional tax τ on equilibrium total food
expenditures mij . While the second constraint is clearly not implementable in practice
(we will consider implementability in section 5.3 below), it captures the assumption that
the planner does not want to change the distribution of total food consumption other
than shifting total food consumption up or down for all households.
The combined constraints on the planner’s problem can be represented by replacing
the aggregate resource constraint (8) by an resource constraint and a “planner budget
constraint” for each location j,∑i∈j
[κH (yij)CH,ij + κU (yij)CU,ij ] =∑i∈j
(1− τj)m (yij) (10)
KH (yj) +KU (yj) =∑i∈j
τjm (yij) (11)
where the planner chooses τj for all locations j in addition to the consumption allocation.
The constrained efficient allocation subject to these constraints, corresponds to the first-
best allocation with appropriate Pareto weights to prevent redistribution. The additional
constraints do not affect efficiency condition (9) for the relative consumption of healthy
and unhealthy food.
5.2 Effect of price distortions on diet
Figure 8 shows the relative consumption of fruit and vegetables in the first-best allo-
cation, as well as in the equilibrium allocation and in the data. For comparison, we
have also plotted the recommended intake of fruit and vegetables of at least 40% (U.S.
22
Department of Health and Human Services and U.S. Department of Agriculture (2020)).
Since the Pareto weights do not affect the relative consumption of healthy and unhealthy
food, the first-best allocation with equal Pareto weights is identical to the solution of
the planner problem without redistribution.
Figure 8: Relative Consumption - First best
30 40 50 60 70 80 90 100
Average County Income (in 1000$)
-3.1
-3
-2.9
-2.8
-2.7
-2.6
-2.5
Cons. F
V / C
ons. O
ther
food
Model
First Best
Data
optimal diet
Note: This figure displays the relative consumption of fruit and vegetable, against county income, in
the data (purple dotted line), in the estimated model (solid blue line), and in the first-best efficient
allocation (dashed black line). The dashed red line represents the intake of fruit and vegetables
recommended by U.S. Department of Health and Human Services and U.S. Department of Agriculture
(2020). Each point is a 5-percentile bin.
Comparing the first-best to the equilibrium allocation, consumption of fruit and veg-
etables is 14 to 15.5% lower than efficient due to price distortions due to supply factors.
This effect is roughly constant for households of all income levels, and across counties
of different average income as well, and there is very little effect of price distortions on
dietary inequality. This explains why the difference-in-differences analysis in Allcott,
Diamond, Dube, Handbury, Rahkovsky, and Schnell (2019), which compared poorer to
richer households, reveals very little role for prices.
The effect of price distortions on diets is large. On average these distortions are
responsible for about one third of the gap between the actual and recommended intakes
of fruit and vegetables, ranging from almost a quarter for the poorest households to
almost the entire gap for the richest 5% of households, with the rest being due to
preferences. It is important to note that the role of supply factors we estimated is a
lower bound. The only inefficiency we consider is the effect of fixed costs in the supply
of healthy food. Other externalities that are likely to be important, most notably the
fact that consumers may not fully take into account the health effects on their diet,12
12This may be due to an ‘internality’, boundedly rational consumers discounting the long-term effectsof their dietary choices because of the difficulty of self-control (O’Donoghue and Rabin (2006), Haavioand Kotakorpi (2011)) or to moral hazard if consumers are insured for their health care costs.
23
are not included in our model and their effect will therefore be attributed to preferences.
In addition, preference are very broadly defined in our model and may themselves be
endogenous and possibly inefficient. This would be the case, for instance, if consumers
are uninformed (or ill-informed) about the health effect of their diet or influenced by
advertising or marketing by the food industry. In this paper, we deliberately focus on
a narrow set of supply factors, in order to make the case that even if we restrict the
analysis to price distortions, for which we find strong evidence from observable data,
there is a case for government intervention in the food environment. We will have nothing
to say about the effect of interventions such as information and education campaigns,
improved food labelling or restrictions in advertising, even though in reality these may
be important aspects of a government strategy to encourage healthy diets.
Figure 9: First Best
(a) Welfare
0 50 100 150
Individual Income (in 1000$)
-15
-10
-5
0
5
10
15
20
25
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
We
lfa
re (
Co
ns E
q)
(pe
rce
nt)
Average Welfare Gain: FB = 0.17
Average Welfare Gain: No Redr = 0.081
Average Welfare Gain: FB = 0.17
Average Welfare Gain: No Redr = 0.081
Equal weights (left)
No Redr. weights (right)
(b) Consump Other food
0 50 100 150
Individual Income (in 1000$)
7.45
7.5
7.55
7.6
7.65
7.7
7.75
7.8
7.85
C O
ther
food
Model
Equal weights
No Redr. weights
Data
Note: The left panel plots welfare gains/losses against household income in the first-best allocation with respect
to the equilibrium, in percentage consumption-equivalent terms. The dotted black line (left y-axis) corresponds
to the first best allocation with equal Pareto weights, while the dashed green line (right y-axis) corresponds to the
first-best allocation with Pareto weights that imply no redistribution. The right panel displays the consumption
of other food in the these two first-best allocations together with the estimated level (solid blue line) and the
data (purple circled line). Each point is a 5-percentile bin.
Figure 9 show the difference between the first-best and the equilibrium allocation in
terms of consumers’ utility and consumption of “other food” (all food except for fruit
and vegetables). With equal Pareto weights, the planner redistributes consumption from
richer to poorer households, so that poorer households are much better off in the efficient
allocation (utility of the poorest 5% of households is higher in the first-best allocation by
a consumption equivalent of 20%), but richer households are made worse off, despite the
improvement in their diet. This picture is dramatically different when we choose Pareto
weights to avoid redistribution. In this case, presented in figure 9, richer households
are unambiguously better off in the first-best allocation, by a consumption equivalent
of slightly over 0.25%, because of their improved diet. However, poorer households
suffer a welfare loss from being taxed for the fixed costs of supplying healthy food that
they would not have chosen to consume in equilibrium. This is particularly true for poor
24
households living in rich counties, where the fixed costs of supplying fruit and vegetables
are higher.
5.3 An uncontroversial tax and subsidy policy
Which households are winners and which ones are losers from a policy that eliminates
price distortions due to fixed costs in the supply of fuit and vegetables depends on how
the planner funds this policy, in particular on how much consumption is redistributed
across households. This raises the question whether it is possible to design a policy
to eliminate distortions that makes all households better off. It is worth noting that
in the no-redistribution case, the average consumer in all counties is strictly better off
in the first-best allocation, as illustrated in figure 10, which plots welfare gains and
losses against county-level rather than household-level income. Within counties some
consumers are made worse off, because for them the utility loss from the higher tax on
consumption exceeds the utility gain from a better diet. We now explore whether it is
possible to compensate these losers through a different tax policy.
Figure 10: Welfare - Location Income
30 40 50 60 70 80 90 100
Average County Income (in 1000$)
-1
-0.5
0
0.5
1
1.5
Welfare
(C
ons E
q)
(perc
ent)
0.076
0.078
0.08
0.082
0.084
0.086
0.088
0.09
0.092
Welfare
(C
ons E
q)
(perc
ent)
Equal weights (left)
No Redr. weights(right)
Note: The left panel plots welfare gains/losses against county income in the first-best allocation with
respect to the equilibrium, in percentage consumption-equivalent terms. The dotted black line (left
y-axis) corresponds to the first-best allocation with equal Pareto weights, while the dashed green line
(right y-axis) corresponds to the first-best allocation with Pareto weights that imply no redistribution.
Consider a household-specific tax rate τij instead of a location-specific tax τj in
equations (10) and (11). We look for an income-dependent tax rate τij = τj (yij) that
generates the same welfare gain for all households within a county. Figure 11 shows that
this efficient allocation is indeed a compensated Pareto improvement over the equilibrium
allocation. The welfare gains from moving to the efficient allocation under this policy are
almost constant for all households at 0.1% of consumption. There are small differences
between households that live in different locations, but since we already showed that the
25
average household in each county improves its welfare, see figure 10, now that the welfare
gains are equalized within counties it must be that welfare improves for all households.
Figure 11: Pareto Improvement
(a) Welfare
0 50 100 150
Individual Income (in 1000$)
0.075
0.08
0.085
0.09
0.095
0.1
0.105
0.11
We
lfa
re C
on
s E
q)
(pe
rce
nt)
Average Welfare Gain: Pareto Impr= 0.102
(b) Consump Other food
0 50 100 150
Individual Income (in 1000$)
7.5
7.55
7.6
7.65
7.7
7.75
7.8
7.85
C O
ther
food
Model
Pareto improv
Data
Note: The left panel present welfare gains/losses against household income in the Pareto-improving
consumption allocation with respect to the equilibrium, in percentage consumption-equivalent terms.
The right panel displays the consumption of other food in the Pareto-improving allocation (orange
dashed diamond line) together with the estimated level (solid blue line) and the data (purple circled
line). Each point is a 5-percentile bin.
Under the Pareto-improving efficient allocation, fruit and vegetable purchases are
subsidized, and households are taxed for other food consumption at a rate that increases
with their income and varies with the location where they live, see figure 12. The
administration costs of this optimal policy would be very high, but it is possible to
approximate it with a simple rule that is feasible to implement in reality. This simple
rule would fund the subsidy on fruit and vegetables through the income tax, which
-different from a sales tax- allows to tax households progressively.
6 Conclusions
In this paper, we explore the question to what extent dietary choices are influenced by
the “environment” as opposed to preferences. This question is relevant, because if diets
are mostly determined by external factors, then there is a stronger case for government
intervention. We aim to establish a lower bound on the role of the environment, and
narrowly focus on distortions in the relative price for healthy versus unhealthy food.
Other possible distortions, for instance because consumers to not (fully) internalize the
health effects of their diet, are outside the scope of this paper.
We find that relative prices of healthy food are severely distorted, and fruit and
vegetables are 40% higher than is efficient. We also show that these price distortions
lead to inefficient underconsumption of fruit and vegetables by 15%, about a third
of the gap between actual and recommended intakes. In contrast, the effect of price
distortions on differences in diet across households of different income levels, is limited,
26
Figure 12: Tax Rates
(a) Individual Income
0 50 100 150
Individual Income (in 1000$)
0.012
0.014
0.016
0.018
0.02
0.022
0.024
tax r
ate
(b) Location Income
30 40 50 60 70 80 90 100
Average County Income (in 1000$)
0.012
0.014
0.016
0.018
0.02
0.022
0.024
0.026
0.028
0.03
tax r
ate
Note: This figure displays the implied tax rates τ in the Pareto-improving allocation as a function ofhousehold income, left panel, and county income, right panel.
which explains the contrast between our findings and those of recent previous studies.
While we use the same data, our approach differs from these previous studies in that
our results are based on estimates of a structural model of dietary choice, which allows
us to estimate aggregate effects and do welfare and counterfactual analysis.
Starting from the consumption allocation a benevolent social planner would choose,
we evaluate different policies to improve diets. We show that it is theoretically possible
to achieve efficient relative price levels without making any consumer worse off. We also
show that a simple tax and subsidy policy, which could be implemented in practice, gets
very close to replicating this first-best allocation, and we document its welfare gain and
distributional effects.
References
Algert, S. J., A. Agrawal, and D. S. Lewis (2006). Disparities in Access to Fresh
Produce in Low-Income Neighborhoods in Los Angeles. American Journal of Pre-
ventive Medicine 30 (5), 365–370.
Allcott, H., R. Diamond, J.-P. Dube, J. Handbury, I. Rahkovsky, and M. Schnell
(2019). Food Deserts and the Causes of Nutritional Inequality. The Quarterly
Journal of Economics 134 (4), 1793–1844.
Altomonte, C., A. Barattieri, and S. Basu (2015). Average-cost pricing: Some evidence
and implications. European Economic Review 79 (C), 281–296.
Amano-Patino, N. (2020). Nutritional Inequality: The Role of Prices, Income, and
Preferences. Cambridge Working Paper in Economics 1909, University of Cam-
bridge.
Andreyeva, T., M. W. Long, and K. D. Brownell (2010). The impact of food prices
27
on consumption: a systematic review of research on the price elasticity of demand
for food. American Journal of Public Health 100 (2), 216–222.
BBC (2014). Obesity is the new smoking, says NHS boss in England. BBC News:
Health, 18 September 2014.
Bell, J., V. Rogers, W. Dietz, C. Ogden, C. Schuler, and T. Popovic (2011). CDC
grand rounds: Childhood obesity in the United States. 60, 42–46.
Cawley, J. (2015). An economy of scales: A selective review of obesity’s economic
causes, consequences, and solutions. Journal of Health Economics 43 (C), 244–
268.
Cutler, D. M., E. L. Glaeser, and J. M. Shapiro (2003). Why Have Americans Become
More Obese? Journal of Economic Perspectives 17 (3), 93–118.
Danaei, G., E. L. Ding, D. Mozaffarian, B. Taylor, J. Rehm, C. J. Murray, and M. Ez-
zati (2009). The preventable causes of death in the United States: comparative risk
assessment of dietary, lifestyle, and metabolic risk factors. PLoS Medicine 6 (4),
e1000058.
Finkelstein, E. A., J. G. Trogdon, J. W. Cohen, and W. Dietz (2009). Annual medi-
cal spending attributable to obesity: Payer-and service-specific estimates. Health
Affairs 28 (Supplement 1), w822–w831.
Haavio, M. and K. Kotakorpi (2011). The political economy of sin taxes. European
Economic Review 55 (4), 575–594.
Kelsey, R. (1989). Packaging in Today’s Society (3d ed.). Lancaster, PA: Technomic
Publishing Company.
Komlos, J. and M. Brabec (2011). The trend of BMI values of US adults by deciles,
birth cohorts 1882-1986 stratified by gender and ethnicity. Economics and human
biology 9, 234–50.
Larson, N., M. Story, and M. Nelson (2009). Neighborhood Environments: Dispar-
ities in Access to Healthy Foods in the U.S. American Journal of Preventive
Medicine 36 (1), 74–81.
Lustig, R. (2012). Fat Chance: Beating the Odds against Sugar, Processed Food,
Obesity, and Diseases. New York: Hudson Street Press.
O’Donoghue, T. and M. Rabin (2006). Optimal sin taxes. Journal of Public Eco-
nomics 90 (10-11), 1825–1849.
Ogden, C. L., M. D. Carroll, L. R. Curtin, M. A. McDowell, C. J. Tabak, and K. M.
Flegal (2006). Prevalence of overweight and obesity in the United States, 1999-
2004. JAMA 295 (13), 1549–1555.
Pincock, S. (2011). Boyd Swinburn: combating obesity at the community level. The
Lancet: Perspectives 378 (9793), 761.
28
Powell, L. M., S. Slater, D. Mirtcheva, Y. Bao, and F. J. Chaloupka (2007). Store
Availability and Neighborhood Characteristics in the United States. Preventive
Medicine 44, 189–195.
Schlosser, E. (2013). Fast Food Nation: The Dark Side of the All-American Meal.
New York: Perennial.
Sharkey, J. R., S. Horel, and W. R. Dean (2010). Neighborhood Deprivation, Vehicle
Ownership, and Potential Spatial Access to a Variety of Fruits and Vegetables in
a Large Rural Area in Texas. International Journal of Health Geographics 26 (9),
1–27.
Sturm, R. (2005). Childhood obesity - what we can learn from existing data on societal
trends, part 2. Preventing Chronic Disease 2 (2), A20.
Townshend, T. and A. Lake (2017). Obesogenic environments: current evidence of
the built and food environments. Perspectives in Public Health 137 (1), 38–44.
U.S. Department of Health and Human Services and U.S. Department of Agriculture
(2020). Guidelines for Americans 2015-2020 (8th ed.).
Wendt, M. and J. E. Todd (2011). The Effect of Food and Beverage Prices on Chil-
dren’s Weights. Economic Research Report 134705, United States Department of
Agriculture, Economic Research Service.
29
top related